An explainable multi-task deep learning framework for crash severity prediction using multi-source data.

Area of Science:

• Traffic safety analysis
• Machine learning in transportation
• Road accident prediction modeling

Background:

• Traffic accidents present significant global challenges, leading to injuries, fatalities, and economic losses.
• Existing research often focuses on single prediction tasks, overlooking property damage and inter-task relationships.
• Neural networks in traffic safety are limited by interpretability issues and data complexities.

Purpose of the Study:

• To propose an interpretable multi-task deep neural network (MT-DNN) framework for comprehensive crash severity prediction.
• To integrate enhanced deep learning with post-hoc explanation methods for causal insights.
• To address multiple prediction targets (fatalities, severe injuries, property damage) and identify key contributing factors.

Main Methods:

• Developed an interpretable multi-task learning framework (Adv MT-DNN) integrating deep neural networks and post-hoc explanation techniques.
• Employed SHAP-based methods for feature importance rankings and interaction analysis.
• Validated the framework using four-year multi-source traffic data from China (2018-2021).

Main Results:

• The Adv MT-DNN framework demonstrated significant improvements in prediction accuracy over baseline models.
• Identified and ranked the top 8 critical factors influencing crash severity, including blood alcohol content and collision type.
• Confirmed statistically significant associations between identified factors and crash severity through nonparametric estimation.

Conclusions:

• The proposed framework effectively bridges the gap between predictive performance and model interpretability in traffic safety.
• Provides engineering-relevant insights and a robust foundation for data-driven road safety policies.
• Offers valuable contributions to developing intelligent transportation systems, especially in complex traffic environments.